Notes from seminar.[ref]These are my own notes as taken during the seminar and may not accurately reflect everything. I merely confirm with the speaker that my notes are publicly visible. My own commentary/thoughts indicated in square brackets. [/ref] Robert speaks to an utterly packed room. Predicting future range of species: 4 topics. 1 cross-validation, 2 niches, 3 unit (sp?), 4 mechanism. Presence-absence predictions from machine learning, maxent.
Opening: AUC for measure of fit- lousy stat!
Cross-validation: Training data vs test data. So can compare models, but doesn’t mean anything in absolute terms. Prediction by distance to known observations of sp occurance works better than bioclim and about as well as maxent- seems maxent just captures spatial autocorrelation mostly, and not ecology/niche!
compare niche predictions based on historical range vs current range- very different. [You live where humans let you much more than where climate lets you. Want to forecast potential niche/climate determinants of niche, but these aren’t often the most informative predictors seized upon by the machine learning].
units of analysis. (Brad’s salamanders?) little evidence that ranges are climate driven.
Mechanistic models- hot topic, but no guarantee of better predictions. Just fitting a story (time management model) isn’t really different. Food & cold are limiting- “Gorillas were driven up these mnts and they don’t like it there. They want global warming” How about one-tailed limits: rather than having dist around optimum conditions, just have single-sided limits (min H2O but no max).
Ends with Levins three model types.
[my question: Is the result- all about people, not climate- due to focus on big mammals and crops? ]